Follow RSS

GUIDES / Office

Turn Meeting Notes into Action Items — a Faithfulness Test You Can Repeat

Two synthetic transcripts (EN/ZH) tested against pre-registered truth: owners, vague deadlines, rejected suggestions, cancelled tasks. Every result, including the two fixes a human still made.

Office · Published 2026-11-11 · YYSI Editorial · 中文版

The short answer

Paste a text transcript into the meeting notes generator, press the run button (1 of your 3 site-wide free trials), and you get a summary, a decisions list, action items with owner and due fields, and open questions — editable in place and downloadable as Markdown. It is text-only: there is no audio upload or transcription in this first version, so get your transcript from your recording app or a transcription service first. Two things it promises not to do: it does not guess owners or deadlines the transcript never states (they render as "not stated"), and it does not send anything anywhere — your text goes through YYSI's server to the AI provider, nothing else.

Warning：A transcript is the tool's only source. If a person's name or a date was never said out loud, the tool must leave that field empty — and in our tests it did. That is the feature that makes human review fast: every filled field is worth checking, every empty field is a real gap in the meeting, not a hidden guess.

What we tested (and how you can repeat it)

We wrote two original synthetic transcripts — an English product sync and a Chinese store weekly — and pre-registered the ground truth before running anything: four action items each with different owner/deadline situations (explicit owner + exact date, explicit owner + vague date, no owner at all, a confirm-by task), one cancelled task, one rejected suggestion, one open question, and a number that must survive verbatim. The transcripts and truth files are downloadable below, so you can rerun the same inputs and compare.

Faithfulness checkEN transcriptZH transcript
4 pre-registered action items found with correct owners4/44/4
Exact dates kept (13 / 18 March; 3月14 / 3月18)kept as 2026-03-13, 2026-03-18kept as 2026-03-14, 2026-03-18
Unassigned task left with no owner (not guessed)yes — "not assigned to a specific person"yes — owner left empty, also raised as an open question
Vague deadline not converted to a fake dateyes — left empty (see review notes)yes — left empty (see review notes)
Rejected suggestion not listed as a decisioncorrectly framed as "withdrew his proposal"correctly kept out; the decision to keep the vendor listed instead
Cancelled task not listed as an action itemcorrectcorrect
Open question capturedyesyes
Numbers verbatim (12,480 CNY / 2,340 元)yesyes
Fabricated people/dates/numbersnone foundnone found

Both runs came back with zero fabricated owners, dates or numbers, and both correctly separated the rejected suggestion from real decisions. n=2 synthetic meetings — this is a faithfulness check you can repeat, not a benchmark.

The two things a human still had to fix

The tool's own review workflow is: read the output, edit inline, copy. Our pre-registered review found exactly two issues per language, both small and both fixed in the corrected files below:

  1. Vague deadlines get dropped, not kept. The transcript said "sometime next week" / "下周找时间" — the tool left the due field empty rather than writing "next week". Nothing was invented, but the vague information was lost. We restored "next week" / "下周" by hand. If your meetings lean on vague deadlines, expect to type those back in.
  2. One reported-numbers line was labeled a decision. "Q1 spend so far is 12,480 CNY, no corrections" / "上月损耗 2,340 元…无异议通过" appeared in the decisions list. It is faithful to the facts but generous about the framing — nobody decided anything about the spend. We re-labeled it as reported information.
  3. Two grey cases behaved acceptably: a wish with no decision ("it would be nice if somebody eventually refreshed the FAQ") was listed without owner or due, with wording that keeps it a nice-to-have; and a passing remark with a name in it ("小林顺手换一下海报") became an owner-tagged item without a deadline.
Note：Everything else was the tool's untouched output. The corrected files below contain only those two edits per language, marked with a comment at the top — diff them against the raw files to see exactly what a human added.

Steps in the tool

  1. Get a transcript. Record your meeting as usual, then produce text with your recording app's transcription or any transcription service. The tool itself takes text only — no audio, no files. Paste-ready format: one line per speaker turn, Name: what they said.
  2. Paste and run. Open the meeting notes generator, sign in, paste the transcript (up to 30,000 characters; at least 50), keep Output language: Same as transcript (or force English/中文), and press the run button — it consumes 1 of your 3 site-wide free trials.
  3. Review against the transcript. Check three things in order: every owner against who actually spoke, every date against what was actually said, and the decisions list against phrases like "we decided" versus "someone suggested". Everything is editable in place — click into any field and type.
  4. Copy or download. Copy the notes as text, or download Markdown (yysi-meeting-notes.md) with sections for summary, decisions, action items and open questions. Results stay available for 24 hours, then expire.

Limits, cost and privacy

  • Input: text only, up to 30,000 characters per task, minimum 50. No audio upload or transcription in this version.
  • Output: capped at 4,096 tokens — roughly 20 decisions + 20 action items + 10 open questions with realistic sentence lengths. A day-long workshop may need to be split.
  • Owners and due dates render as "not stated" when the transcript never says them. Dates may be normalized to ISO form (our test's "13 March 2026" came back as 2026-03-13) — same day, different format.
  • Cost: each run consumes 1 of your 3 site-wide free trials (never reset). After that, usage packs apply — the pricing page has current numbers. Copying, editing and re-downloading a finished result cost nothing extra.
  • Privacy: the transcript is sent through YYSI's server to the AI provider (DeepSeek). YYSI stores character counts, token usage and the generated notes; notes are kept for 24 hours, then become unavailable. Nothing is emailed or posted anywhere.
  • Known boundaries: complex tables in transcripts, heavy crosstalk, and speaker names never spoken aloud are all weak spots — check those lines by hand.

Who this is for (and who should skip it)

  • Good fit: regular team syncs where a human types or paste-transcribes notes and wants the action items split out with owners and dates — weekly ops, project stand-ups, client-call debriefs.
  • Skip it if: you need one-click recording-to-minutes (not supported here — the SaaS bots that join your calls do that, and they cost monthly fees), or your meetings are mostly free-form brainstorm with no tasks at all.
  • Verification-heavy teams (finance, legal, ops with money on the line): use it as a first draft only — our test still needed the two human fixes above, and amounts deserve line-by-line checking.

Try it with our material

Everything below is the exact material from the runs: the two synthetic transcripts (all names fictional, no real meetings), the pre-registered truth files, the raw tool outputs in Markdown form, and the human-corrected versions whose only changes are the two review fixes described above. Paste a transcript into the tool and compare your result with ours — outputs are AI-generated and can vary slightly between runs, which is exactly why the truth file, not a single expected string, is the reference.

Capability statements follow the tool page (checked 2026-11-11). The tested engine version is 62e9d82d857f, byte-identical to the deployed page at test time.

Sources & verification

Sources checked All test numbers come from YYSI's own runs on 2026-11-11: two production meeting-notes-generator tasks (task ids t_0004392cf9a76cb27058f973 EN, t_eeca26758c0c83a47137a8b8 ZH) on a standard test account (3 site-wide trials; 2 consumed; no account or ledger changes; live engine hash 62e9d82d857f matched the workspace build byte for byte). Because the operator's browser-to-Cloudflare connection was degraded that day, the runs were submitted to the same production task endpoint the page's button posts to, with byte-identical transcript input and identical tool/locale/funding parameters; the page UI render path was not exercised and this is disclosed in the guide. Scoring rules and ground truth were pre-registered before the runs (test-design.zh.md in the private run directory); scorer v1 had two over-strict checks (full-name attribution, decision-to-keep wording) that were corrected to implement the registered intent — both versions' differences are documented in the run report.

Read our editorial standards or send corrections to [email protected].